TobyYang7/shadow-interface-evolution-trajectories

Shadow Interface Evolution Trajectories

OSWorld V2 evaluation trajectories from Shadow Interface Evolution, and the evolved shadow-interface source trees that produced them.

Archive	Evolution experiment	Variant	Trajectories	Full-credit accuracy	Mean partial score
prime25v15r4_if-006_trajectories.zip	prime25v15r4	if-006	68	17/68 (25.0%)	0.6240
opus5xover25v15_if-010_trajectories.zip	opus5xover25v15	if-010	68	16/68 (23.5%)	0.5834

Each archive contains only claude_trajectory.jsonl files, preserving the original evaluation batch and task directory structure. Separate screenshot files, result files, and task-cache artifacts are not included.

Scores use all 68 held-out tasks as the denominator, counting infrastructure failures as zero. Both runs have infrastructure-imputed zeros for tasks 002, 018, and 024. Trajectory presence does not imply a successful or complete task execution.

The if-006 archive combines the original 33-task batch and the remaining 35-task batch. The if-010 archive is from the crossover experiment (opus5xover25v15), evaluated in run 20260906T174130Z.

opus5xover25v15 seed

opus5xover25v15_seed_trajectories.zip contains 40 claude_trajectory.jsonl files from the independently evaluated opus5xover25v15-seed, preserving the original task directory structure. It covers 25 training tasks and 15 validation tasks; this is not a held-out 68-task baseline. Separate result and screenshot files are not included.

Across these 40 tasks, full-credit accuracy is 4/40 (10.0%) and mean partial score is 0.3135557204. These scores are from the seed records in shadow_evolution_osworld_v2_opus5xover25v15.db and are not directly comparable to the held-out scores above. The recorded run uses Qwen3.8-27B through OpenRouter/novita, a 1M context window, and a two-hour per-task time limit.

prime25v15r4 seed

prime25v15r4_seed_trajectories.zip contains 40 claude_trajectory.jsonl files from prime25v15r4-seed, preserving the original task directory structure. It covers 25 training tasks and 15 validation tasks; this is not a held-out 68-task baseline. Separate result and screenshot files are not included.

Across these 40 tasks, full-credit accuracy is 7/40 (17.5%) and mean partial score is 0.3890158997. These scores are from the seed records in shadow_evolution_osworld_v2_prime25v15r4.db and are not directly comparable to the held-out scores above. This seed was evaluated independently of opus5xover25v15-seed; their results and trajectories are distinct.

The recorded run uses Qwen3.8-27B through OpenRouter/novita, an effective 1M context window, and no per-task wall-clock limit. Although its saved interface configuration declares context_window_tokens: 131072, that field was not wired into Prime during this evaluation, so the model catalog window applied. Other step and runtime termination conditions may still apply.

qwen_synthetic30v15 (seed and if-001)

qwen_synthetic30v15_seed_trajectories.zip and qwen_synthetic30v15_if-001_trajectories.zip contain 38 trajectory files each from qwen_synthetic30v15-seed and its one mutation qwen_synthetic30v15-if-001, preserving the original task directory structure. Alongside the 36 claude_trajectory.jsonl files, each archive carries the two retry/phase trajectories that those runs produced (claude_trajectory_attempt_2.jsonl, claude_trajectory_phase_1.jsonl). Separate result and screenshot files and task-cache artifacts are not included.

This experiment uses a different split from the archives above: 30 synthetic training tasks (s_001–s_030) plus the same 15 validation tasks. It is not a held-out 68-task baseline.

Archive	Variant	Trajectories	Val full-credit	Val mean partial
qwen_synthetic30v15_seed_trajectories.zip	seed	36 (21 train + 15 val)	1/15 (6.7%)	0.2349933333
qwen_synthetic30v15_if-001_trajectories.zip	if-001	36 (21 train + 15 val)	2/15 (13.3%)	0.3243400000

Only the validation column is a score. The 30 synthetic training tasks have no verifiers; their recorded score is a dummy constant zero, labelled in each result.json as UNGRADED: constant zero; use trajectories, not scores. A combined 36-task accuracy would therefore be meaningless, and none is given. Validation figures come from the variant records in shadow_evolution_osworld_v2_qwen_synthetic30v15.db.

Nine training tasks are missing from both archives: s_003, s_010, s_011, s_013, s_016, s_020, s_022, s_027, s_028. These are exactly the nine synthetic tasks that drive a web application, and all nine failed environment setup on every attempt in both rounds, producing no trajectory at all. The cause was a defect in the imported task package, since fixed: the web task configs never launched Chrome, so setup could not reach it over CDP. Both variants in this experiment were therefore evolved and evaluated against 21 of 30 training trajectories, all of them desktop-only. The validation gain from seed to if-001 should not be read as evidence about browser tasks.

The recorded run uses Qwen3.8-27B for both the builder and the task agent through OpenRouter/novita with fallbacks disabled, the Prime Agent harness, an interface-declared context window, 300 agent steps and a two-hour wall-clock limit per task, and a linear mutation-only route with latest-parent selection.

qwen_syn30v15_fixed (seed + five evolution steps, with software repositories)

This is the corrected rerun of the qwen_synthetic30v15 experiment above. The task-package defect that starved the earlier run of its nine web tasks is fixed, so every variant here was evolved against the full 30 training trajectories (bar six individual task failures noted below). It runs five evolution steps instead of one, with crossover enabled at cadence 2 and tournament parent selection, which produces a branching lineage rather than a linear chain.

It is also the first entry in this dataset to publish the evolved software repositories alongside the trajectories. For each variant, qwen_syn30v15_fixed_<variant>_software_repository.zip contains the complete shadow_interface/ source tree at that variant's git tag — the artifact the builder actually evolved, and the thing the agent imports at task time. qwen_syn30v15_fixed/manifest.json records each variant's git tag and commit, parent, operation, step, validation score, and file counts.

Archive pair	Variant	Operation	Parent	Trajectories	Source files	Val full-credit	Val mean partial
qwen_syn30v15_fixed_seed_*	seed	seed	—	45 (30 train + 15 val)	10	1/15 (6.7%)	0.2350
qwen_syn30v15_fixed_if-001_*	if-001	mutation	seed	42 (29 train + 13 val)	14	0/15 (0.0%)	0.1747
qwen_syn30v15_fixed_if-002_*	if-002	crossover	seed	45 (30 train + 15 val)	14	3/15 (20.0%)	0.2850
qwen_syn30v15_fixed_if-003_*	if-003	mutation	if-002	45 (30 train + 15 val)	15	1/15 (6.7%)	0.3390
qwen_syn30v15_fixed_if-004_*	if-004	crossover	if-002	44 (30 train + 14 val)	15	2/15 (13.3%)	0.2267
qwen_syn30v15_fixed_if-005_*	if-005	mutation	if-003	44 (30 train + 14 val)	21	1/15 (6.7%)	0.1879

Only the validation columns are scores. As above, the 30 synthetic training tasks have no verifiers and record a dummy constant zero labelled UNGRADED: constant zero; use trajectories, not scores. Validation figures come from the variant records in shadow_evolution_osworld_v2_qwen_syn30v15_fixed.db.

The lineage branches: seed -> if-001 and seed -> if-002, then if-002 -> if-003 and if-002 -> if-004, then if-003 -> if-005. The peak is if-003 at 0.3390 mean partial, a 44% relative gain over the seed; if-002 has the highest full-credit count. The two steps after if-003 both regress, so this run gives no evidence that further steps would help.

Six trajectories are missing, and none of them is the old Chrome fault. Across all 270 task runs there were zero ECONNREFUSED failures, zero browser UTC failures, and zero tracebacks. The gaps are: s_003, 046 and 072 in if-001; 072 in if-004; 007 and 072 in if-005. Five of the six are a transient outage of the teamchat.site.hku.icu task website, which recovered during the run; 007 in if-005 is a user-simulator failure. Task 072 is the seed's only full-credit validation task, and losing it in three separate variants depresses those variants' validation scores relative to the seed.

The recorded run uses Qwen3.8-27B for both the builder and the task agent through OpenRouter/novita with fallbacks disabled, the Prime Agent harness, an interface-declared context window, 300 agent steps and a two-hour wall-clock limit per task, 40 concurrent VMs, and a deterministic scheduler with crossover cadence 2. Total API cost across the six variants was $396.13, of which $4.60 was the builder and $391.53 the task agent.

GitLab personal access tokens that a task agent minted inside a task VM's own GitLab instance have been redacted from the if-004 archive. Those VMs were destroyed at the end of the run; the redaction is a precaution, not a response to a live credential.